Decomposer: Learning to Decompile Symbolic Music (Like MIDI) to Programs (arxiv.org)

Researchers have introduced Decomposer, a groundbreaking framework for decompiling symbolic music, notably MIDI, into executable programs in the Strudel music programming language. This innovative approach addresses the complex inverse problem of translating musical performances back into high-level instructions, a task complicated by the scarcity of aligned data between MIDI and code. Decomposer employs a two-stage methodology: first, it utilizes a synthetic corpus, Strudel-Synth, to fine-tune the model with paired MIDI and Strudel programs; then, it applies reinforcement learning on unpaired MIDI to enhance both the accuracy of MIDI reconstruction and the readability of the generated code. The significance of Decomposer lies in its potential to bridge the gap between musical performance and programming, making music creation more accessible to coders and composers alike. The model achieves superior performance in MIDI reconstruction compared to existing closed-source language models, showcasing a marked improvement in code readability and diversity compared to traditional heuristic converters. This advancement not only elevates the state of music programming but also opens new avenues for creative expression and computational musicology, positioning Decomposer as a pivotal tool for the AI and music technology communities.